Why value and process matter more than technology

Full session transcript

THIS IS AN AUTOMATED TRANSCRIPT

Noz Urbina: Hello everybody. My name is Noz Urbina and welcome back to the Omnichannel X podcast. I have with me today here Lasse, who I have been corresponding with on LinkedIn and I’ve been wanting to have on this podcast for the past few years.

Lasse is the host of The Only Constant podcast. He’s also the AI lead at Basico. And like myself, he’s quite interested in demystifying complex technological trends and fostering open dialogue about all these opportunities and challenges presented by emerging technologies.

I like Lasse because his interview style is deep, it’s dynamic. He likes to just have an open chat and is open for any conversation about the future of work and technology. He’s also a historian who used to do poetry and has one of LinkedIn’s best post to laugh-out-loud ratios. In my experience he’s actually quite funny, so I recommend you follow him just for that, to make this sometimes turbulent world a little bit lighter. So Lasse, welcome to the show.

Lasse Rindom: Thank you so much Noz. Thank you for that introduction. It’s great to be here and I always look forward to speaking to you. And I remember when we did our episode, we also agreed that we were gonna keep on talking. Right. And maybe, oh, let’s do another one. At one point you were so happy. I think you felt like this was the best day of your life. And then I was like, no, no, no. You only get one.

Noz: Right. Yeah. No. I went into a deep depression after that. That’s why it’s been so long.

Lasse: So really happy to be here with you today again. No, so I love it.

Noz: It is an honor. So about your podcast, you’ve been doing The Only Constant podcast for, oh, a year now, over a year.

Lasse: Yes, one and a half years exactly since November 2023.

Noz: Right. And you are much more productive than I am. I think you podcast every week. So that’s quite a pace. We just switched to this seasonal rhythm where we’re doing 10 episodes a year because weekly podcasts, I’m not the guy. But I do think we are gonna focus on making sure we do 10 quality episodes a year so people can come back when the season begins. How has your perspective on change and AI evolved since starting a podcast called The Only Constant?

Lasse: Well, I think, first of all, about me making an episode every week – I actually decided last year that I wanted to do one every second week. But then I ended up having enough interesting guests that they just get booked up. So they started being once a week again. These are one hour episodes, very unstructured, deep conversations with thought leaders. So it takes up some mental energy as well.

But I think the reason that I did it was because my curiosity is just so deep. I want to know what’s going on. It’s called The Only Constant because you want to find this change thing. I used to do poetry because I was looking for some kind of truth. I used to do history because I was looking for some kind of truth. I’m still looking for truth, but I think the podcast is my meditation saying that there is no single truth. So I keep asking everyone all the time, what is the only constant? Because I know there is no only constant. But it’s still the curiosity that drives me.

And how has my perspective on change evolved since I launched it? I think it has deepened and widened. That’s the boring answer. I think actually this meditation on change being something where there’s no single point of truth, as I started out with, is something that just keeps resonating in the episodes. But I always try to find new angles. I want to hear new things.

The challenge is when you’ve heard things being said so many times, you might sometimes have episodes where you think, oh, nothing really new came up there. But then you have listeners saying, oh, this was an awesome episode. It’s just because I’ve done now 67 episodes, so I’ve heard everything before.

But change is nothing if it’s not attached to value. I think that’s just one thing that’s become increasingly clear. I started it off being focused on AI. I still talk a lot about AI on it. But the whole AI thing, I think over time, it’s still the background, but it’s not at the center anymore. The centerpiece, the play we are unfolding, is more value and outcome right now. And I think that’s more interesting. Technology is a tool. In the end, it is a tool. I don’t think anyone wants to say technology is all, it’s all about all the time. So we always end on this value thing in the end instead.

Noz: Well, it’s interesting. We talked a lot about value in the last couple episodes on OmniX and I can remember we talked about truth, and I think that’s one of the reasons we hit it off. The phrase “value over volume” is one that I’m using a lot these days. Because I think it’s so central when we talk about these technology operations – it’s so easy to take a mass production, acceleration, operational efficiency perspective, whereas there’s this huge opportunity to talk about operational quality.

My favorite metaphor, I can’t remember who mentioned it first, but I didn’t really agree with his conclusion. He said AI is like plastic. It allows us to mass produce mediocre knockoff artifacts that clog up our world.

Lasse: And medical equipment that we really need, you know, sterile equipment. You can’t talk about a general purpose technology like that. It’s like saying electricity is just used for crap because it’s in my kids’ toys. But it’s also everywhere else, right? So I think people are missing the point when they say things like that.

One of the things I’m saying all the time at the moment is if you don’t start with outcome and outcome discussion about what you want, then you’re not gonna end up with outcome. And that has nothing to do with AI at all. That’s in everything, no matter what you want to do.

Tom Reiner, who was on my podcast as well and used to be my boss for a brief period at HFS Research, he was once asked about, is RPA dead? And he was like, if it solves the problem, it’s not dead. And I think people really need to understand that it’s all about solving problems in the end. And if a technology solves a problem, it’s a good technology, period. When you go to SMBs, you’ll see that very clearly because they have to take money out of their own pockets. So they’re like, will this solve my problem in a better way? No? Then I won’t buy it. Period.

Noz: There’s a circularity to that argument though. First of all, it’s awesome that you interrupted me and finished my sentence exactly as I was gonna finish it because I reposted that guy’s post and said, are we the ones making the plastic that moves precious materials and medical devices? Or are we making party favors and single-use Halloween decorations? It’s a choice how you choose to apply these technologies and what value and the longevity of that value that you try to get.

But I think there’s a circularity to when people say, will it solve my problem? You have to understand it to be able to answer that question. So I think there’s a certain amount of learning and experimenting you need to do in order to be able to understand how it could solve your problems.

What I’m seeing is mass rollout at every level of the chatbot. When we say AI today, for most people, that’s become almost synonymous with chatbots. ChatGPT has pulled a Google – they are the name in the common consciousness, which I kind of hate because I have a lot of issues with OpenAI as a company, Sam Altman in particular. But I train on ChatGPT in my workshops because that’s what everybody has.

What I’m seeing is people don’t have the mental models, the baseline understanding to be able to think: How can I use this thing? So that’s where we get immediately in discussions like, oh, it’s not accurate like Excel is, it’s not accurate like my database is, therefore it’s useless. Can I use it like a database? No, therefore it’s useless.

So I think that we do have to understand what an AI process looks like and what re-engineering processes for AI looks like before we can answer the question of, is this gonna be useful? Is this gonna solve my problem? Maybe it’s solving problems that you didn’t identify, or problems that you would not have attempted to solve had you not had it.

Lasse: Yeah, but I agree with that. But we’re getting back to something that’s an old problem as well. How do you explain to people what automation can do in general? I think 10 years ago I said that people looked at their applications like they were God – set in stone. That’s the way they work. You can’t change it. The ways of the Lord and SAP are mysterious. But you can change things. You can automate with it. But that’s an abstraction level that people need to understand and which takes time for them to understand.

But let me just give you a number: there’s 127 million daily users of ChatGPT. More than 700 million monthly users of ChatGPT. That’s the web online version on OpenAI’s website. That’s a lot of users. This means this is happening. This is already there. People are using it. If you’re not using it, you can be pretty sure that the junior consultant you’ll hire tomorrow will be using it and he’ll be challenging you because he’s using it. So you’re not gonna necessarily get replaced by him, but you just need to up your game and get started on it.

My point is, if this technology needs to be pushed to start, it’s not a good technology. This does not need a push. It’s gonna come. It’s there, it’s happening. And I don’t think you need to push it. I don’t think you need to go around saying all the positive things about AI right now. I think you just need to say, you need to start working with it, period. It has things you can do obviously. That’s why everyone’s investing so much money in it. It’s not a fad. It’s the thing. Get over it, move on. I’m not asking you about your opinion if you believe in it or not. I’m just saying you have to relate to it some way or another. And I don’t care how, because this is already happening.

Noz: I don’t disagree. I agree with everything you’re saying. I think I’m making a somewhat parallel or maybe subtler point. I’m remembering the rise of social business. And I remember being laughed out of the room when we were first talking about social media as a business tool. Remember that? Like the first few years and it was, no, it’s a kid’s thing. So it may be rising everywhere, but we’re a serious enterprise.

Lasse: People think about it as something that just makes funny videos, makes funny images, funny songs. They know it can check an email and that’s like, oh, that’s really cool. But it is funny. Except I think it’s different.

I work a lot in the CFO and COO space. That’s where I’m predominantly spending my time. And you are in the marketing space. So there, it’s the content and the channels and all that. It’s like completely different. But looking at your transactional tabular data and saying, where can I use this language model that just made a funny poem about financial planning and analysis? I don’t know where I can use that. It seems silly.

Noz: The difference I’m making is that with social media, you didn’t have CEOs having it on every desk in the company saying, go get on social. That never happened with social. So there was, from the top, there was an association of silliness, which prevented this kind of aggressive push that’s happening from top down.

But it was just an example. My point is that what I’ve seen is when a technology is sufficiently new and different, people often run in with their old templates in their head and their old processes and they try to apply them to technology. It doesn’t work. They get worse results and then they get burned and progress gets stalled. That’s what I’m trying to refer to. AI is coming like a tsunami – there’s no argument about that. But in the world of business and how do we actually get productive, do we kind of flatten the curve of bad first experiences? That’s what I’m talking about – facilitating this change.

Lasse: But the problem is that a lot of people have started this the same way that it started with every other change. Everyone’s got an AI budget for trying it out. Enterprises have been trying it out and they set up a department to do AI and try something out and made a chatbot to chat with our data or something like that. Tons of them out there. Millions of them.

And then in the end, they were like, okay, this is play around money, fun money. And then a year later they’re looking at, oh, no one’s using it. What’s wrong? Why aren’t we getting any ROI from this? And you’re like, but it was play around money. You wanted to just play with it and see what it could do, but now you’re suddenly expecting it to have had a real business impact. So now you’re disappointed with it, but you didn’t expect something from the beginning.

So I think back to what I said before – we started the wrong place. We didn’t start by asking what do we want? What is the outcome we want to achieve? And I think what we need to do right now, more than ever, and maybe we should have also done that in 2024 and 2023, is to expect a KPI from whatever we do with AI. We do that with everything else. We expect some kind of measure so we can say we got a return. But we haven’t done that with AI for so long. We’ve just been like, just because we’re doing AI, it’s valuable.

But let’s say, okay, we wanted to review all of these invoices. Good, okay. Did you review all of them? How many errors were there? What is the uptime? It’s simple to add some KPIs, but just measure it. Say, no, it didn’t do what it’s supposed to do. Can we make it better? Yes, we can. But just adding those numbers to it, I think that’s simple. And then we’ll get to a place where we say, oh, here it works. Oh, here it doesn’t work.

But saying, here’s a bunch of data we have, let’s make a chatbot to chat with it – you have no clue what will make this a success. You have no clue what’s going on. You need to have a measure to say, this didn’t work or this did work. Instead of saying yes or no to AI, we need to say what are the pitfalls? What are the things it can’t do? What are the things we can expect it to do in the future? Much more interesting to discuss that. But it doesn’t work well in hyped up messages about AI will ruin everything in the world. It becomes much more complex because guess what? The world is complex.

I think I said it in a show with Andreas Welch last year that it’s crawl, walk, run, right? And we got the tech, but we didn’t get the change. Because change is us changing.

Noz: That’s, I think – don’t know if it might have been on your podcast – I said that AI’s adoption will be… the limiting factor will be human change.

Lasse: Yeah. But there’s also just the whole brownfield-greenfield thing. Everyone’s talking about the first billion dollar AI company. I don’t think so. Most companies get entrenched. I think people are so blindly looking at, okay, I can make a whole law firm now with this technology. Yeah, but if you could do that, then it’ll become commodity and worthless. So you won’t make a billion dollar company with AI, because if AI could do it, it’s not worth a billion dollars. Period.

Noz: Wait, wait, wait…

Lasse: No, no, no. Let me just… Now you distracted me. I had another good point. Crap.

Noz: I am sure it’ll come back.

Lasse: Yeah, it will. Okay. So what I’m saying… I think everyone’s talking about a billion dollar company. Everyone’s talking about it in our LinkedIn filter bubble. But when I go out of that and I hold an open training and I say, what are your AI use cases, people? And I go to my clients and I ask them, what are your AI use cases? They blink at me and they go, spell check and grammar.

But that’s because that’s exactly back to my point. My point is that most companies are not this greenfield thing. Even the greenfield will not work.

Noz: Define greenfield, brownfield for…

Lasse: Greenfield is just, there’s nothing there. You can build something new, you can do whatever you want. You can build on any structure you want. You can say it’s AI driven from the beginning. There’s no structures in place. Brownfield means that something has already been built there. It’s like when you build architecture in the world – there’s nothing there, it’s just a greenfield. You build a house, perfect. But if there’s something there already, you need to tear it down. You need to integrate into the sewage systems and electricity fields and everything. It becomes much more complex.

And here’s my key point: I keep on repeating to people, the world is brownfield. Everything out there is brownfield. Your company exists. It’s a weird thing to have to tell people that your company exists, right? It’s there. It’s having a value prop, delivering value to the customer, to the world. It exists out there and you need to change it.

That’s actually what I’m going around saying. I’m just going around saying to people, there’s something out there. You wanna have impact, then you have to realize there’s something you’re impacting. If a meteor has impact, it destroys things. Things are moving around. What is it you’re moving around? What is it you are expecting to move around? What are the people, the processes, the values, the customers, the stakeholders? What is it you wanna move? That’s what you need to discuss right now. Not just talk about necessarily AI, but figure out how do we move these things and what is the value that drives these things?

Not just saying that it’s stupid that we have these silos. Why do we have the silos? What are they producing? What is the purpose of them? There’s a purpose behind everything we’re doing. Sometimes – and this is subtle – sometimes the purpose of a thing is to slow things down. We’re talking so much about speed and AI and automation that no one can even fathom that sometimes it’s just a good thing that things are slow, because that means that we can predict what’s going on. That’s the reason why we have speed limits on the roads. There’s a reason why laws take months to get approved because we want things to move slow so no one can just ruin everything. That’s the point of a brownfield – things need to change.

Noz: Don’t we wanna move fast and break stuff?

Lasse: Yeah, we don’t wanna move fast and break stuff. We have things that work. I think we need to respect that more than in an AI craze that things need to change. One of my favorite quotes is from Enzo Torresi. He’s an old enterprise architect. Maybe God made the world in seven days, but he didn’t have an installed legacy base.

Noz: So I think that brings me to another topic. I think we’re violently agreeing, which is always a good thing. That process definition, process analysis, understanding what you’re trying to do, how is it gonna work, how are the people that you have today gonna be part of that process are all the essential questions. But you’ve interviewed…

Lasse: People wanna buy AI and put it somewhere and say, now we have AI. Now we’ve added AI.

Noz: It’s the magic wand thing. I’ve been suffering the magic wand sales pitch my entire career. It’s just every few years, some big thing gets hyped. This just happens to be the biggest one, but it’s the same pitch: No more thinking. No more work. Only happy time.

Lasse: You’ll get tired of winning. But I think executives will think this is just another fad like that. The board is saying it, and the market is saying we need to do it to keep our employees and brand and all that. So now we’ll do it, but we’ll put it over here in this little bucket where we have all the creative people and they’ll be doing something creative with this creative new thing. And then we’ll go around doing our stuff in Excel over here as we used to do. That’s what we do. We just keep on moving. And whenever we meet them on the corner, we’ll be like, hi, that’s the nice guy, that’s the AI guys. But they’re really not contributing anything because it’s just added on. They’re not impacting anything because here we’re busy. We’re doing things, we’re delivering to our clients. This is serious business. It’s fine that you can do AI images here, but…

Noz: That’s not what I’m seeing. I’m seeing corporate-wide rollouts of the CEOs saying to everybody, every department – cleaning staff, HR, cafeteria – how are you gonna use AI? I’m seeing the board pressure to make stuff better. Here’s, I just bought you a magic. Everybody in the company now has a magic wand. I bought you all magic wands. Now go make magic. Show me. And that is a different but also very difficult situation.

Lasse: But I think there’s a different thing to that. You’ll also be hiring the couple of AI managers to be the AI lead, the COE. The COE is also happening at all these companies and then they’re all going to get a license for Copilot or ChatGPT, and then you expect 10 to 20% more productivity. So they’re setting those up. And that’s what I call these things – we put it over here and they’ll be the creative ones. But then you also push everyone saying that now you have access to it. But I don’t think change just happens that way, just say, now you have access to it.

Access is not adoption. Adoption is still that people feel like… The key problem with AI right now – and I think this will actually change on its own, that’s why again, pushing the car to start, this is changing because companies are investing so much in it. But the key problem is that you’re taking your processes and giving it to an AI. You’re taking it out to an AI to make AI productive for you. You just have to have a little workshop with it. So you discuss with it what is my problem? Then it understands it, and then it gives you something valuable. But most people don’t have time to do that. Most people are entrenched in something they’re already doing.

So what we’re waiting for right now is that this is gonna get built into the applications they’re having. That’s not just a Copilot thing. This is built into the ERP systems, the CRM systems, the work order systems, everything people are working on daily. It should just be natively doing things that are productive, not just being a chat. Don’t take your processes to a chatbot, but put the chatbot into your processes. I think that’s what people are waiting for.

Noz: And is it just a chatbot? I agree that it should be baked into our processes. Absolutely. But what’s happening now is this stupid refine this text or make this more clear or change the style. We’re just getting more and more advanced versions of Grammarly. Basically, can I rewrite a piece of text with an LLM in whatever application? And I find those implementations so boring because they don’t actually understand the business context of the task at hand.

Lasse: No, but that’s what I’m talking about. When they become these agents in the systems and it starts working in your processes. Right now, you still have to take your process and your documents and everything and put it into a chatbot. That’s how people work with it. But in the future, these chatbots are gonna be doing things in your systems, not as chatbots, but they’re gonna be interacting with data. They’re gonna be basically behind the scenes writing code to manipulate your data and everything that’s going on.

And I think that’s what people are always missing. It’s a language model, but language code is also language. So you can manipulate everything. Digital can be manipulated by an AI. And people don’t realize these things. So there’s just… but that also means that then suddenly the LLM – let’s say that instead of saying the chatbot – the LLM will come into your processes and start working in your processes. And that’s where we’ll have the big benefits of it. Not by just saying, here’s an LLM, let’s see where you can take this.

Noz: Okay, so let’s make this more concrete. Because I was scared when I had you on the show that we’re insiders. I don’t wanna have an insider discussion or only an insider discussion. So you’ve interviewed a ton of people. What patterns have you noticed in how successful organizations adapt? Let’s put some examples and some concreteness on this.

Lasse: I think the issue is it is changing as we speak, but not a lot of companies have been really successful. I think that’s the big challenge. People have been asking me for a long time where are the best AI agents and who’s been… where, what are the best applications of AI? You tell me. I haven’t seen them. I’ve seen RPA agents – a completely different thing. I haven’t seen these big AI agents yet at work, and I know that also companies like SAP, like Business Central who I’ve had on my podcast as well, they are spending years building these AI agents and making them work. It’s not something that happens in a couple of months.

Noz: Does it have to be agents though, or have you seen anyone with any kind of success with these new types of generative technologies?

Lasse: I’ve seen scenarios where you’ve had AI determine where a spare part should go based on where you have the largest penalty if operations are down. So it determines the site that needs the spare part more than the others, and then suddenly we’re getting some real value.

You can also see AI starting to use complex contracts to review things that are actually working and happening in work orders. So if you’re doing things on this site, but then suddenly AI can read the contract and say, does it make sense that we’re doing this? Or do we have some issues that no normal human would ever be able to do on site, on spot, immediately. So these things are adding a lot of value to how we can work.

But it takes time for people to understand that this could happen because this also moves across silos. This is something that we typically put in the law department, the legal department, and then suddenly it’s in the operations department. And bridging those two things also means we have to discuss ownership. Who owns the things and all that. So things are happening.

The first adopters have obviously been marketing and content creation, and that’s where everyone’s been thinking, oh, that’s perfect for that. I think that’s gonna change. I think it’s gonna be the opposite. I think you’re gonna have more impact on the transactional data stuff than actually on content and marketing, because I think we’re gonna realize shortly that our customers are our most important people and perhaps we should be very direct and human when we are interacting with them. But all this transactional data, who gives a crap? And these contract reviews, no one cares about, no one wants to do it. Let’s put an AI to do it. It makes total sense, but it’s just not been the obvious one because we thought generative – make stuff.

Noz: I hate the word generative because for me it’s about restructuring, translating, and interpreting data much, much, much more than generating new crap for the world.

Lasse: That’s very interesting because you said earlier that I am in marketing, although that is one of our big pillars. But what I’m usually doing in marketing is looking at process and looking at not how do we make the stuff. I’m not a copy specialist or a tone and voice person.

Noz: I know you’ve been in this market where it’s been exploding like crazy. Like everyone’s generating blah, blah, blah.

Lasse: No, no, no. And I’m not insulted. That’s a hundred percent true. In my world, it’s more like we have numbers. We don’t need a language model. That’s been my world.

Noz: Right. No, but I think the overlap between where I mainly work and what I bring to marketing and what you’re doing is the look at how do we look at the data we do have, what understanding we do have, make sure that that understanding is getting where it needs to go. And how do we use this technology not to generate stuff necessarily, but to better understand the experiences we’re trying to deliver.

So although I’m in marketing, that’s what I’m doing in marketing. Why is our marketing working or not working? Or if I’m in technical documentation or experience design or product design, why is the thing that we’re making succeeding or not succeeding? And how can we use AI to understand what we’re putting out into the world, what we know about the world, how we map that together, and how we make strategies and plans.

I’m also advocating against too much generation. I am pro-generation when it’s the synthesis of a whole bunch of other things. So if I have a language model, I can ask you a very complex question. Like I’m having a dinner with these eight people, I’m thinking about doing a fusion of these two cuisines, and these people have these dietary restrictions. You can just say all this and that will be formulated into a complex query into a bunch of databases in the backend, which will bring you back a bunch of answers.

So the AI is a little thin language layer translating between good structured data that is gonna keep the AI truthful, keep it from hallucinating. And when it is generating, it is just generating a wrapper or a nice way to present what backend systems have provided. That’s the kind of generation that I like – when we are not allowing these models to run wild as they were when they were first being rolled out. We were giving them minimal control, minimal actual guardrails, and then they would go screw up and they would create bad content or lie or hallucinate.

So I am really encouraging people to jump into generation. I’m encouraging them to say, how can you use this as much upstream in your internal processes for understanding your content and organizing your content and your data, and mapping that to your experience so that you work so much better? Every human being in the company is now working so much more effectively that the company performs better, not how can we fire five people and replace them with a machine that generates their work product.

Lasse: Then it becomes all about speed. I think the problem is that we are focusing so much about speed. I actually heard a podcast yesterday about how that’s actually a bias in our human brain that we tend to look at speed all the time. The way we ourselves look at the world is by inferring what’s going on all the time, because that’s the speedy way of doing it. We’re actually not wired for figuring out how it really works. That’s why you had this blue and white, white and gold dress. Remember that one? That’s because there’s different ways the speed of our brain infers what’s going on with that dress. That’s because of a bias in your brain that just says, this inference makes sense to me. Your brain just says, this is for speed. This is what we need.

So we want speed all the time. That’s why people are always looking at this. How can we fire the people? How can we make things more efficient, faster, faster? Instead of saying, how can we make more quality? I think the quality step is the difficult one, and that’s the one we need to train ourselves on. That’s also why I think I’ve tried to repeat outcome and value. Let’s look at what is the quality we want to do and have that discussion. And that’s actually a very difficult, uncomfortable discussion. People think it isn’t, but it is.

But to what you said just before, I’m starting to always say that processing – every kind of processing – is metadata production. Whatever you do, you take some iron in and you redo it and add some metadata. Now you say it’s a frying pan. The frying pan is a metadata thing on top of that piece of iron you have. And everyone will be like, yeah, I can definitely totally see this is a frying pan.

But the point is – and maybe that’s an extreme case – but every time you take an invoice you say, okay, what is the metadata? This one has to go to this approval. It has to go on this account. It’s attached to this client. You add metadata to it all the time. And if there’s one thing these models can do, not flawlessly but to a very large degree because they’ve been trained on so much data, they’re able to add metadata to almost anything you give them.

Try taking an object – this is a practical thing. Try taking anything, you can put there a picture of a thing or document and say: Dear AI, what metadata, what interpretations would you put on this thing, whatever I’ve given you? Then you’ll realize very quickly what this can actually do for you in your processing, because your processing is about this.

And that also means that – here’s my big overall way I think whenever I ideate with any client at any point about AI – and I know it’s impossible because the only constant is that entropy will continue. The universe will end in a big void of nothing. Energy just dispersed.

Noz: Your poetry is coming out.

Lasse: Yeah, it is coming out. But I’m always saying: is this solution gonna reduce the entropy or increase the entropy in this process? So what I’m basically saying is, is this adding to the chaos?

Noz: So translate entropy in this context for those who are not familiar.

Lasse: Is this adding to the chaos? Taking something that is already structured data and making it unstructured in a gibberish blah blah post? Or is it taking that gibberish blah blah post and turning that into structured, tabular data that I can move in my process? Am I structuring the world more or less? That’s what every company does. That’s what every city does. That’s what everything we ever do – that is civilization at its core. We take things that are chaotic. Titans, the Greek titans, and we kill them, and then we build things that are orderly with roads that are straight, that are boxes, and everything’s just in order and it’s illuminated.

That’s what we do. That’s what companies do. They have an edge of customers that want different things, but we say you can only get these 20 things. We structure them into whatever we can produce. We reduce the entropy of the universe that way. That is the point. We want to do that. We can’t do it, but that is what we do with civilization, and that’s the same thing that AI can do. It can help us structure things that are unstructured.

Noz: So I think in terms of successes, what we’re seeing – and we have been seeing way before the generative wave – was auto categorization or auto tagging. How can I take all this information which I have and I can’t find it? I don’t know. I can’t link it. I can’t give it context.

So I think two big things that are rising is that auto tagging was already huge and already useful, and now it’s becoming much more mainstream. And what you can do with those tags once you have them is now more valuable because then you can tag something, and then when you do want to generate, then you have so much more context around what it is you’re working on.

And the other thing is context management. I’ve been saying for 20 years – I used to call myself and I still do call myself, depending on the circles I’m in, a content strategist. And we always joked about we’re actually context strategists. And so then now there’s this big movement from prompt engineering to context engineering. How do we make sure that a human or an automatic system has the right data, the right structures, the right meta information, the right context to do their job?

If you come to the prompt, you always have to come and give it the prompt and then it’s gonna go do its job. But how have you prepared it to do its job? And when these AIs are gonna be released on the internet and your little paragraph of a request is a tiny drop in the bucket compared to all of this information it’s looking at, how do you make sure that what you get is actually focused and scoped and narrowed down and looking at the sources that you actually trust?

So building context, I think is one of the applications of what you’re talking about. We wanna make more order of this big, disorderly mass of content that we’re constantly producing. And then from that order, we want to choose which parts of it we feed into our processes. What do we want? If I’ve got a complex product with many versions, do I make sure that when I ask a question, you’re only looking at the versions of the things that I own? If I’m in a particular hospital with particular equipment in a particular situation, how do you make sure that you’re providing me the instructions and the spares? Or if I am making purchase decisions and there’s all sorts of terminology and things that I care about when I’m making a purchase decision, how do I make sure that you know enough about my context, so you bring me the right information? I think that’s the next big thing – ordering the world and then using that order to build the right context for people and AI.

Lasse: Or accepting that the world is not ordered, but you build an AI that is able to actually order it. And I think what you just said made me think that one of the big differences when we now have these foundational models is that we were used to saying that technology or something we built – we start building this part and then we increase it, we add onto it. But here we’ve gotten a full model. So we have a really, really big model and now we need to decrease it to be purposeful. So we need to narrow it down. It’s the other way around.

Because if you take full unstructured reality and you give it to a full unstructured AI model, you’re not really increasing anything. Either side has to be restricted a lot. So either you restrict the thing you give it very much, or you restrict it a lot. So you add roles to it. And I think that’s what people also need to understand.

I wrote a mock post a couple of weeks back where I said that the budget for AI is binary. It’s either $1 billion or $20. That’s what you wanna pay. You don’t wanna pay anything in between. Because you see it working – I’m giving it to my ChatGPT. See, it works. Yeah. But you want it not just to work now, you want it to work again and again, in the same way. And that means you have to restrict this really big monster of an AI model to be very concrete.

And that means you have to understand how you give it enough focus to be narrow and not say anything about Lord of the Rings when you’re asking about your business processes. And then also give it enough context to know what you’re trying to work on. So you have to expand it on the context, but narrow it on the focus.

And then you have to make sure also that your semantic distance between the concepts you’re asking about are not too great. Because if you’re asking about some things that are completely disconnected, it will make weird routes and find a way to connect them. But it will always make something different to you.

And then you have to ask it multiple times. That’s the last thing. That’s how you make a good prompt. You have to ask it over and over. So if you ask one AI agent, it will give you an answer. The best example is this: What should I use to edit my photos? It will say Photoshop 60% of the time. Okay, good. But then it’ll say Paint also 5% of the time. But if you ask another one, “Was that a good answer?” and it said Paint, they will say, no, it’s probably more Photoshop. And then you reduce your hallucinations by 70%. People don’t know that. But just adding this chain, just programming it, spending a little more than $20, not $1 billion, but just a little money to make sure you get the… that’s what people need to do. But people are like, oh, but ChatGPT does it almost for me. Yeah. But almost is not good enough for anyone. You need to have something that’s reliable when you start.

Noz: Okay, so I agree and disagree. I agree that any intelligence can use a second pair of eyes, whether they’re digital or human. So if an AI gives an output, it’s much more efficient to have another AI check that output than it is to try to build a super AI who gets it right the first time. So that I totally a hundred percent agree with. I don’t agree that less than a hundred percent reliability is not useful because…

Lasse: No, no, no. But no, that’s not what I meant. High reliability, you want to get high reliability. Not just like, oh, 80% is okay, let’s just put this to do all of our… One in five is wrong is not a good thing. So you need to make sure… and you’ll get that if you just ask it a simple prompt and you’ll get something that’s sometimes different and it moves on in different routes. You need to make sure you test it. Unlimited capabilities means unlimited testing, so you need to find a way to narrow it down just a little bit.

Noz: Agreed. Well, this brings us back to context engineering. How can we make it so that when I’m doing my work, I don’t want to be prompt engineering. How can I make it so that the thing has enough context that I’ve set up my AI properly so that when I put in a simple prompt, it has enough context to not require me to explain the whole world to it to give me the answers I want?

Lasse: Either you have to pre-prompt it or you’ll have to prompt it. And if you don’t pre-prompt it, someone will have to pre-prompt it. And that’s what I said as well. When it comes into your application suites, these application providers will pre-prompt it.

And it gets me to some of the really biggest thoughts I’ve had – I don’t know if it’s because of the podcast, but during the podcast – is that the future will be a lot about owning the process. So software vendors will have to own the process. They’ll have to define the process because you can’t make an agent work in that process without that.

I think people are not realizing that digital beings like agents or AI agents will only move in a digital space. And if your digital space is composable with unclear unmatched APIs and bad data structure and bad master data, they will not be able to work. Ambiguity kills the AI cat. Most of companies’ digital infrastructure is ambiguous because we have human operators at the center of processes and applications and software have typically been functional add-ons outside that you just use whenever you need it. But right now you’re putting technology at the center. If you have this composable, ambiguous setup, it won’t work. It can’t navigate.

So I think the dialectic dial is moving back towards a little bit more monolithic – not necessarily in the UI monolithic sense, but in just, this is a defined process. This is how we’re doing it, this is what it means. This is a data layer. This is how you… this is what a customer is, this is what a churn customer is. All these things have to be strictly defined so AI can move around.

Noz: I’ve been trying to order people’s content and data world so that people can move around for 25 years, so I’m totally there with you for that. I tell people that I’m in the “I told you so” years of my career. Because we’ve been saying, you cannot have all this stuff loosely defined in 20 different systems and just have it work out.

Lasse: You can do that with humans to sort of stop the gaps around. But if you want AI agents, you can’t do that.

Noz: Yeah. So hopefully that will be the thing that finally breaks the back of this.

Lasse: All these platform providers – a lot of companies are just providing platforms to do whatever you wanna do, but right now, they need to start defining the process or they can’t do gen AI. And the companies that will define a good process for you or define what your data means, they will win in the race at least. So if agents go to that scale that everyone’s thinking they’re going to, you have to be able to, as a software provider, define the process. You’ve heard about this, maybe this service as a software? That means that you are outsourcing your services to a software provider. Then that software provider better take responsibility for your process and define it as well. Most of them won’t. Most of them are just like, you do whatever you wanna do here, and there’s a disconnect there.

Noz: So okay, so we gotta wrap this up. So I want to come down to… You’ve been podcasting now for a year and a half. You talked to a lot of great people. I’ll be interested in… You can repeat some of the things we’ve already said, but what would be your kind of top takeaways from a year and a half? So let’s pull out a bit and go, what are those moments that made you go, “Ha!”

Lasse: That is the one thing you asked me to prepare for. So I actually wrote… I think you asked me for top five, right? And I actually wrote seven because that’s the way I am. I always do more, over deliver and annoy people a little bit. But that’s the way it is.

But I think Yann Kun said that the only constant is that you always have to innovate on top of a change commoditized baseline. Things have changed. Everyone has AI today. Everyone has it. It’s not something that will make you stand out. You’re not different because you have it. It will not be your secret sauce because it’s everyone’s sauce.

Noz: Oh, so what is a commodity has changed and you have to innovate on top of it.

Lasse: Gen AI is a commodity. Period. Understand that. You have to innovate on top of that, and you have to think about how you do that. And I think that’s what’s difficult. That also means that suddenly you’re thinking about, okay, how is the ecosystem gonna change? Because it’s not just gen AI. We have to focus on the downstream changes that come from gen AI. So people always wanna have these blinders. You have to remove those and think how does it look beyond that? What’s going to happen beyond just the AI thing? Because AI is not interesting. AI is interesting right now, but in a moment, in a heartbeat, it won’t be. Value will be – change roles, change value props, all that.

And that’s what you need to innovate on top of that. And I love that he said that. He’s been part of the development team of the Google Assistant, been CEO for Google in Switzerland. So he knows what he’s talking about as well in technology.

Noz: You said it was Yann Kun.

Lasse: Yann Kun. I can’t remember the number on my show.

Noz: Ah, okay. No, it’s interesting because Yann LeCun has just come out a couple months ago saying, I’m no longer interested in LLMs.

Lasse: Exactly.

Noz: Which I love. He’s moving on to what he calls world models, which for me would be knowledge graphs or domain models where you define the world.

Lasse: Define the world. So he’s moving on to where I think he should have been. So okay, so that’s two people who…

But that’s one thing. And then the second one is that regulation doesn’t mean you don’t have… deregulation doesn’t mean that you don’t have responsibility. This was Kate O’Neill who said that, and I really love that one. Because I think everyone’s talking about this deregulate so we can move faster. But I think people need to understand, if you deregulate things, it just means that every single company has to work out their own way of interpreting what’s good and what’s not good. And the race is not about having the most AI, it’s about having the most adopted AI. And adoption sometimes needs regulation, period.

Noz: That’s a… I love that. I love that quote.

Lasse: Yes. I love it too. It just makes me… well, we’re gonna talk about that more on the Truth Collapse Podcast, but makes me very stressed.

And then you have Robert Feldt who is building human in the loop and driving systems for Zenuity. He’s the head of R&D for them. And I spoke to him about human in the loop in general because he’s been working with that for 25 years in cars. So how do you add… I think I love that conversation. One of my favorite ones, because it was just so new for me and all the things. He was like, I’ve been working with this forever and everyone in automation business is like, oh, human in the loop, we’re saying it, but we don’t really know what it means.

But he talked about the human crumple zone. You know, when a car crashes, there’s a crumple zone in front that protects the human. But in complex systems, the human becomes the crumple zone for the complex system. When you have aviation, you have the autopilot, then you always blame the pilot. But the pilot has no clue to understand what’s going on with the sensors or anything in the plane when it crashes. But they get the responsibility. Even though… because the human has to protect the complex system. The same thing happened with Three Mile Island. All those things.

And are we building human crumple zones now with AI where we’re just blaming humans for… they have no clue. They can’t know because now it’s not even… you can’t even check what made the LLM come up with this idea? It just made it. It’s so complex. It’s so overly complex, so you’re adding it on top of an already complex enterprise architecture. So the human will be the one saying, oh, we just have a human controlling and checking if everything’s all right. Please. I don’t want that job. Somebody wants the human to strangle when something goes wrong. You can’t strangle the AI.

Noz: Exactly.

Lasse: And I just love that concept, the human crumple zone. And then I have Carl Friston as well who said that if it takes more than five minutes to understand, it’s probably worth your time. And that became sort of my motto for my podcast. It’s super crazy difficult podcast, but if it takes more than five minutes, it’s worth your time. Because it will make you more intelligent, more engaged, more like an agent, a human agent.

And he said that to understand this, to innovate, they have to do more, listen to more jazz, he called it. You have to listen to something within a structure that’s confined, but still a little bit chaotic. So you don’t know, you test things out. You go bungee jumping, you don’t base jump off a mountain, but you do bungee jumping. And I think there’s just… that mindset just really resonated with me for how to be an intelligent person, intelligent company, do things the right way. You have to have room for this unstructured exploration as well.

Then versus composable, I just spoke about before – something I thought a lot about after having spoken to SAP and Business Central especially, and also others.

And then I really love the quote from Lorian Pratt, who invented decision intelligence and machine learning transfer, where she said on my podcast, the Wright Brothers might have invented the airplanes, but they did not invent the airline. And then I always expand and say they didn’t invent the airport either. So where are we in the world today with gen AI? Well, we need to invent these airlines. How are we gonna commercialize this? What value are we gonna get from this? And then the airports – how are we gonna structure it? How are we gonna govern it? How are we gonna maintain it over time? And that’s just such a simple image of where we are and what we need to think about. People have airplanes flying all around. How do we make them productive?

Noz: Nice. Beautiful. I like that you saved that one for the last.

Lasse: Yeah. It’s a good one, right?

Noz: Ah, that’s a good one. Very nice. Yeah. Dr. Lorian Pratt, one of my favorites. She’s super smart. Loved her.

So I thought you said Kate O’Neill.

Lasse: Did I say Kate O’Neill said the last one? I meant…

Noz: Yeah.

Lasse: Dr. Lorian Pratt.

Noz: I thought Kate O’Neill got two.

Lasse: Crap. That’s the problem when you’re just talking live here, right? Sometimes you mix up a word. Kate O’Neill was the one with regulation. Lorian Pratt said the one with the airplanes. I hope so.

Noz: No problem. I’m sure you’ll be forgiven by our listeners.

Lasse: I hope so.

Noz: All right. So thank you so much, Lasse. And we’re gonna jump over to the Truth Collapse podcast and we have already some carryovers that we’re gonna talk about. I think that one about the regulation and… I’m interested, I love the airplane metaphor, but I might have some interesting counter thoughts.

Lasse: Love looking forward to it.

Noz: Okay. Fantastic. See you in a couple minutes. And everybody else, you can join us over on truthcollapse.com/podcast if you wanna keep listening. But otherwise, like, subscribe, tell your friends and please give us your feedback through LinkedIn. You can hit me or Lasse. There’s not a lot of Lasses and Nozes, so I think it should be pretty easy to find us. Yeah, let us know your thoughts, your questions, who we should have on the show next, and any other feedback that you have. Thank you Lasse for joining us, and we’ll see you next time.

Lasse: Thank you for having me.